danelcsb/localagent-dispatch-data

LocalAgent Dispatch Data

Synthetic data for training/evaluating a generable tool-dispatch model over a 50-tool surface (route head → dense selector → pointer-copy). A static snapshot of the deterministic generators in LocalAgent (src/localagent/data/). Train/eval are disjoint in both phrasing and slot values. Companion model + demo: danelcsb/localagent-tiny-30m-byte · Space.

Configs

config	rows (train/eval)	what it is
paraphrase	1000 / 1000	many natural phrasings per tool (all 50 tools), single-turn
contextual	230 / 230	referent-conditioned: same instruction → different tool by referent ("Open status.host.io"→open_url, "Open 'Chrome'"→open_app, "Open docs/intro.md"→read_file)
scenarios_single	60 / 40	clarify / abstain (over-trigger negatives, no tool) / parallel (≥2 calls)
scenarios_episodes	60 / 60	multi-turn episodes: workflow / chained / error-recovery
freeform	86 / 45	hand-written natural queries — train (86) + out-of-distribution eval (45)

Schema

Single-turn (paraphrase, contextual, scenarios_single):

{"prompt": "...", "kind": "tool|text", "category": "...", "tool": "<name|>",
 "target": "{\"arguments\":{...},\"name\":\"...\"}", "calls": [{...}]?}

• kind="text" (clarify/abstain) → no tool; target is the natural-language reply.
• calls present for parallel turns (≥2 tool calls). Tool-call arg values are literal substrings of prompt (so a copy/pointer mechanism can ground them).

Episodes (scenarios_episodes): {"category": "...", "turns": [{"role", "content", "tool_calls": [{"name","arguments"}], "tool_response"}]}. Copy-args in follow-up calls appear verbatim in a prior tool_response.

freeform: {"prompt": "...", "tool": "<gold tool>"}.

Usage

from datasets import load_dataset
ds = load_dataset("danelcsb/localagent-dispatch-data", "paraphrase", split="train")

Regenerate deterministically from source: python scripts/export_dataset.py in the LocalAgent repo.